Two sides of the site
AI for energy in data centres means software that learns from operating data and changes a setting so the site uses less power. It works on one of two sides, and each side is graded by its own number.
On the building side, it reads temperatures, flows and weather, and tunes cooling setpoints, fan speeds, pumps and chillers around the load. Its results show up in the facility overhead and in PUE.
On the load side, it reads what the computing equipment is doing and adjusts what the hardware draws while the work runs. Its results show up in IT power and in output per unit of energy, such as tokens per watt.
The two sides are not rivals. Less heat from the load leaves less for the cooling to remove.
What load-side AI listens to
A load-side system needs a live view of each accelerator. The usual source is the platform's own management interface, which reports power, clock, temperature and utilisation for every GPU.
Some systems also read the throughput counter the serving stack already publishes, so they can see the work alongside the watts. Counters like these are enough to judge the draw against the work.
What it has to recognise
The hard part is not reading a number. It is working out what kind of work each accelerator is doing, and noticing when that changes.
The right target depends on the answer, and the answer changes many times a day. The guide to reducing GPU power consumption shows why one fixed setting cannot keep up.
What it decides, and what reaches the hardware
A load-side system turns what it recognises into a target, then writes that target through an interface the hardware vendor publishes. For a GPU, the practical lever is the clock the accelerator runs at, which largely sets how much power it draws.
Every target should pass safety limits before it reaches the part. A clamp keeps it inside a safe range, and a thermal limit can hand control back to the hardware's own defaults.
Stopping the software should release control, and the vendor's own reset commands should return the device to its defaults.
What it leaves alone
Power control works on the hardware, not on the work. The workload, the container, the scheduler, the model and its weights, the prompts, the tokens, the network and the storage can all stay outside it.
Deciding where work runs belongs to the scheduler. A load-side power system acts on the accelerators it has been given, where the work already runs.
What to ask a supplier
Ask which interface it writes through, what it writes and how often. Ask exactly which counters it reads.
Ask what privilege it holds on the host, and how it comes off. Ask what happens to each accelerator when the service stops.
Ask whose instrument measured the result, over what window and against what comparison. Ask whether throughput and latency held on the same run. The guide to energy efficiency in data centres covers how to read the figure itself.
Questions people ask
- How does AI reduce energy use in data centres?
- On the building side it tunes cooling and distribution around the load. On the load side it reads what the computing equipment is doing and adjusts what the hardware draws.
- What does GPU power management software change?
- Typically an operating target such as the clock each accelerator runs at, written through the hardware vendor's own management interface, inside safety and thermal limits.
Where ATHLAZ fits
ADAPT is load-side AI for GPU fleets
ADAPT, AI-Driven Adaptive Power Technology, is a per-GPU software runtime that installs beside the driver. Its loop has four steps: listen, recognise, decide, adjust.
It listens to each accelerator's power, clock, temperature and utilisation counters. It recognises the kind of work by reading the draw in the frequency domain rather than as a single number. It decides a correction for each component of the draw rather than one for the whole signal.
It adjusts by sending one locked clock target per accelerator, per control tick, through NVIDIA's own management interface. Each target passes a safety governor and a thermal manager first, so a reviewer has one interface to check.
Measured: up to 21% less GPU die power on NVIDIA H100 NVL, over a continuous 48-hour window, with managed and baseline arms under an equal power cap. Derived from the same run: 0% throughput change.